Dynamic residual deep learning with photoelectrically regulated neurons for immunological classification

Dynamic deep learning is considered to simulate the nonlinear memory process of the human brain during long-term potentiation and long-term depression. Here, we propose a photoelectrically modulated synaptic transistor based on MXenes that adjusts the nonlinearity and asymmetry by mixing controllable pulses. According to the advantage of residual deep learning, the rule of dynamic learning is thus elaborately developed to improve the accuracy of a highly homologous database (colorimetric enzyme-linked immunosorbent assay [c-ELISA]) from 80.9% to 87.2% and realize the fast convergence. Besides, mixed stimulation also remarkably shortens the iterative update time to 11.6 s as a result of the photoelectric effect accelerating the relaxation of ion migration. Finally, we extend the dynamic learning strategy to long short-term memory (LSTM) and standard datasets (Cifar10 and Cifar100), which well proves the strong robustness of dynamic learning. This work paves the way toward potential synaptic bionic retina for computer-aided detection in immunology.